Supervise an agent while it is on the call
Autonomous voice agents send the money and change the records. The moment that matters is the beat before they act. Watchtower puts a human there, with the full picture and the trigger, and lets them intervene without the caller ever feeling the seam.
Role
Product design and design engineering
Surface
Real time supervision console
Built with
Hand coded HTML, a live state machine
Data
Synthetic
01 · The problem
takeover onlyA takeover toggle is not supervision
Vida's agents act on their own across voice, text, and payments, in verticals where the wrong move is expensive: a charge over the limit, a disclosure without consent, a promise the business cannot keep. The only human control on the surface today is a switch that lets a person jump into the chat.
A switch answers can a human step in. It does not answer the questions that decide the outcome of a live call.
The questions that decide the call
What is the agent about to do, before it does it. What is it allowed to touch on this call. And can I stop it in the two seconds I have, without the caller hearing a seam.
Confirmation prompts do not scale to an agent that runs for minutes across many steps. By the time a human reads a prompt, the money already moved.
02 · The thesis
Move the decision to the beat before it acts
An agent proposes. A human disposes. The whole path stays legible and reversible. Watchtower makes three things visible at once and controllable in real time: what the agent is saying in the live transcript, what it may do as scope, and what it is about to do as a pending action. The supervisor does not read a log after the fact. They hold the trigger.
The Act Window
When an agent queues a consequential action, Watchtower lifts an Act Window a beat before it fires, and puts the action on a short hold that fails closed. Silence never authorizes a charge.
03 · Signature craft
The Act Window: catch the action in the seconds before it fires
The window states the action in plain terms, measures it against scope, and offers four moves: approve it as is, edit the parameters so it lands inside scope, hold the agent so the caller hears a natural stall instead of dead air, or take over the call as a human. If the action is out of scope, approving is a deliberate, logged override, not a default. If the supervisor does nothing, the hold expires and the action is withheld while the agent escalates.
Worked example 1 · payment · scope breach · synthetic
fully interactiveFully interactive. A caller asks to pay $340 against a $200 payment limit. The Act Window lifts, auto held. Approve the override, edit down to $200, hold Ava, or take over. Do nothing and it fails closed. Built to match the Vantage system. Data is synthetic. If the embed is blank, open it in a new tab.
04 · Trust signals
Scope on the left of the decision, a sealed audit on the right
Scope
flips to breach in real timeScope is not a settings page buried three clicks away. It sits beside the transcript as the agent works, so the boundary the agent is pressing against is always in view. When the agent crosses it, the relevant capability flips to a breach state in real time.
Audit
byproduct, not a reportUnderneath, every intervention writes to a tamper evident audit trail with a wall clock timestamp: scope tightened, charge held, override approved, control handed to a human. The record is the byproduct of using the tool, not a report someone remembers to file.
05 · The harder gate
When the caller changes their mind mid call
Scope is a limit a supervisor can override on purpose. Consent is not. In a regulated vertical, some gates belong to the caller, and no amount of supervisor authority should be able to waive them. This second call makes that concrete. A patient asks for a lab result, then stops the agent partway: she is on a line that might be recorded and does not want her medical details read aloud.
The instant consent is withdrawn, the PHI scoped capabilities revoke themselves, and the queued disclosure does not go to a hold, it goes to blocked. There is no approve. The supervisor's real choices are to re request consent, route the result to the patient's secure portal, or take over, and even under a human takeover the PHI gate stays closed. The design says something plainly that most agent tooling avoids: the person supervising the machine is still not the person whose consent it is.
Worked example 2 · healthcare · consent withdrawn · synthetic
hard block, no approveFully interactive. Consent is withdrawn partway through. Watch the PHI capabilities revoke, the disclosure hard block, and choose the safe path: the secure portal keeps the patient served without speaking a word of PHI on a flagged line. Data is synthetic. If the embed is blank, open it in a new tab.
06 · The system
Seven states, and the failure modes that are the point
A supervision surface earns trust in the moments that go wrong, so those were designed first.
01 Watching
nominalAgent nominal, low risk. Calm. No pending action.
02 Elevated
approaching scopeThe agent is approaching a scoped boundary. Amber, before anything is queued.
03 Pending
act window upA consequential action is queued and auto held. The Act Window is up.
04 Held
stall, not silenceThe supervisor paused the agent. The caller hears a stall, not dead air.
05 Takeover
human liveThe human is live, the agent muted, with a handoff marker in the transcript.
06 Handback
one line briefThe agent resumes with a one line brief, so it never contradicts the human.
07 Resolved
audit sealedAction fired or withheld, the call continues or closes, the audit is sealed.
Failure modes, designed first
the point of the toolNo response before the hold expires.
Fails closed. The action is withheld and the agent escalates. The timeout never fires the action.
Out of scope action.
Auto held. It cannot fire without an explicit override, and the override is logged against the supervisor.
Consent withdrawn mid call.
PHI and payment scope revoke immediately. A pending disclosure hard blocks. Not the supervisor's to override.
Takeover mid sentence.
The agent stops at a clause boundary. The caller hears a natural bridge, not a dead cut.
Control link degraded.
Approve and edit disable, since the command cannot be guaranteed to land. Take over stays available through telephony.
Two calls need attention at once.
The higher risk call surfaces and pulses. A pending action on another line cannot be silently missed.
Considered and passed on
looks safe, is notA confirmation prompt on every consequential action
The obvious answer is to make the agent ask a human to confirm before it does anything that matters. It fails at agent speed. An agent that runs for minutes across many steps generates confirmations faster than a person can read them, and a queue of prompts trains the supervisor to approve on reflex, which is worse than no gate at all.
Watchtower inverts it: the agent acts within a scope it was granted, and a human is pulled in only at the boundary, with the action framed, the countdown running, and a default that fails closed. Fewer decisions, each one real.
Explore more
Where this connects
Two neighbors: the observability console this is built to match, and the compliance agent that applies the same accountability.
Vantage, time to insight
The observability console Watchtower is built to match: diagnosis in a read lane, execution handed to a human gated control surface.
View case study
Conformly, EAA compliance agent
The same accountability applied to compliance: an agent whose stated status is exactly what it can stand behind.
View case studyLive supervision for agents that act. Built to be shipped, not just shown.
View all case studiesWatchtower is a concept exploration by David Paterni. Interfaces and data are synthetic.